A large-scale pre-trained RVC model specialized for Japanese pronunciation
Overview
HiFi-GAN / contentvec
Hakumai is a pre-trained model focused on breath (inhale/exhale) to achieve authentic Japanese pronunciation and realistic vocal delivery.
Why Only 32 kHz?
Higher sampling rates such as 48 kHz turned out to be extremely sensitive to handle, with almost no audible benefit.
While they might slightly reduce latency during real-time inference, the difference is negligible.
When fine-tuning for a target speaker, higher rates often caused issues such as reverberation noise and instability.
After considering multiple factors, the model outputs at 32 kHz by design.
If additional high-frequency range is desired, it’s better to expand it afterward using tools like an expander or similar processing.
V2(Stable)
- 147 Speakers(All japanese): 61 Hours
- SR: 32Khz
- Batch 64
- FP32
V5(Professional use)
Superior quality pretrained model
- 225 Speakers(All japanese)
- SR: 32Khz
- Batch 4
- FP32
- Remastered datasets
- LR D/G : 4e-5/1e-4
- Shared layers LR : 2e-6
- Encoder/Flow LR : 1.0e-5
- Vocoder LR : 1.8e-6
- Optimizer/weight decay : AdamW/5e-4
- Mel/KL Decay Start Ratio : 0.58
- Mel/KL Decay End Ratio : 0.9
- Mel Min Scale : 0.45
- KL Min Scale : 0.35
- Freeze foundation model speakers (109)